Kolors

What Kolors is

Kolors is a large text-to-image model built on latent diffusion, released by the Kolors team at Kuaishou on 6 July 2024.^[1] Latent diffusion means the model does not paint pixels one at a time. It learns to gradually remove noise from a compressed internal representation of an image while being guided by a text prompt, and then decodes that representation back into a picture. Working in this compressed space rather than at full resolution keeps training and generation cheaper, which is why the approach is behind many widely used image generators.

Where it came from

Kuaishou announced the model as open source at its WAIC forum on 6 July 2024.^[2] According to the project's documentation, it was trained on billions of text-image pairs and accepts prompts in both Chinese and English, with a 256-token context.^[1] Kuaishou said the model was trained on billions of Chinese-language data points and described it as the most proficient Chinese text-to-image model available.^[2] That Chinese-first emphasis is the project's main stated differentiator. Prompting in Chinese is not simply a translation problem: Chinese text has no spaces between words, so how a prompt is split into tokens and how the text encoder represents it both matter for whether a scene is rendered as asked.

Reported benchmark results

The project published its own comparison set, KolorsPrompts, containing more than 1,000 prompts in 14 categories. On it, 50 expert raters gave Kolors the highest overall satisfaction score (3.59) and the highest visual appeal score (3.99) among the models compared, and its MPS score of 10.3 slightly edged out Midjourney-v6 at 10.2.^[1] Kuaishou also said the model scored 75.23 in an evaluation by the Beijing Academy of Artificial Intelligence, placing second globally among text-to-image models.^[2] The project's README further reports second place on the FlagEval Multimodal Text-to-Image Leaderboard, in a news entry dated 3 July 2024.^[1] These numbers come from the team's own benchmark and from Kuaishou's announcements, so they describe how the developers measured the model rather than an independent consensus.

Licensing and access

Kolors was released with a split licence. The code is under Apache-2.0, and the model weights are fully open for academic research. Commercial use requires registration, and products with more than 300 million monthly active users need a separate licence.^[1] This pattern — permissive for research, conditional for large-scale commercial deployment — became common among open-weight image models, letting researchers and smaller developers work with the model while keeping a negotiation point for the largest platforms.

Ecosystem and follow-ups

The release was followed by a series of integrations and derived models. Diffusers support arrived in July 2024, the README lists ComfyUI support on 9 July 2024 and ModelScope support on 10 July 2024, and the team later published IP-Adapter, ControlNet and inpainting models, along with a Kolors-Virtual-Try-On demo in September 2024.^[1] Diffusers, ComfyUI and ModelScope are the surrounding tooling many practitioners use to run and chain image models, so support for them is what turns a set of weights into something a user can actually put into a workflow. IP-Adapter and ControlNet, in turn, are techniques for steering generation with a reference image or with structural cues such as edges and poses, and inpainting covers editing part of an existing image.

Where it sits

Kolors belongs to the wave of openly released Chinese image models that appeared in 2024, and its distinctive claim is Chinese-language prompt handling at a time when much of the field was benchmarked in English. Its practical footprint comes from three things together: open weights for research, compatibility with the standard open-source tooling, and a family of add-on models for control and editing rather than a single checkpoint.^[1]